1 / 51

KOUROSH MESHGI

Occlusion Aware Particle Filter Tracker to Handle Complex and Persistent Occlusions using Multiple Feature Fusion. KOUROSH MESHGI. PROGRESS REPORT TOPIC. To: Ishii Lab Members, Dr. Shin- ichi Maeda, Dr. Shigeuki Oba, And Prof. Shin Ishii 9 MAY 2014. TRACKING APPLICATIONS. Entertainment.

darren
Download Presentation

KOUROSH MESHGI

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Occlusion Aware Particle Filter Tracker to Handle Complex and Persistent Occlusions using • Multiple Feature Fusion KOUROSHMESHGI PROGRESS REPORT TOPIC To: Ishii Lab Members, Dr. Shin-ichiMaeda, Dr. Shigeuki Oba, And Prof. Shin Ishii 9 MAY 2014

  2. TRACKING APPLICATIONS Entertainment Public Surveillance Robotics Video Indexing Action Recog. Kourosh Meshgi– ISHII LAB - DEC 2013 - Slide 2 MAIN APPLICATIONS

  3. TRACKING CHALLENGES Abrupt Motion Occlusion Illumination Deformation Clutter Varying Scale Kourosh Meshgi– ISHII LAB - DEC 2013 - Slide 3 MAIN CHALLENGES

  4. BAYESIAN TRACKING States: Target Location and Scale • Goal: Define p(Xt|Y1,…,Yt) given p(X1) Observations: Sensory Information Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 4

  5. INTRODUCTION • • • • • • • • • • • • • • • • • • • • • • • • PARTICLE FILTER TR.

  6. INPUT IMAGE Frame: t • RGB Domain Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 6

  7. INPUT DEPTH MAP Frame: t • Depth Domain Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 7

  8. SENSORY INFORMATION Frame: t • Sensory Information Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 8

  9. STATE REPRESENTATION & OBSERVATION MODEL Frame: t • State w (x,y) h Observation Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 9

  10. FEATURES Color Texture • Feature Set Edge Shape Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 10

  11. TEMPLATE Frame: 1 • Template f1 … fj … fM Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 11

  12. PARTICLES INITIALIZATION Frame: 1 • Particles Initialized Overlapped Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 12

  13. MOTION MODEL Frame: t → t + 1 • Motion Model Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 13

  14. FEATURE EXTRACTION Frame: t + 1 • Feature Vectors f1 f2 fM … X1,t+1 X2,t+1 … XN,t+1 Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 14

  15. FEATURE FUSION Frame: t ! • Probability of Observation Each Feature Independence Assumption Kourosh Meshgi– ISHII LAB - MAR 2014- Slide 15

  16. PROB. CALCULATION Frame: t + 1 • Particles • Brighter = More Probable Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 16

  17. TARGET ESTIMATION Frame: t + 1 • Feature Vectors Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 17

  18. MODEL UPDATE Frame: t + 1 • New Model • Model Update Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 18

  19. RESAMPLING Frame: t + 1 • Proportional to Probability Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 19

  20. CHALLENGES • • • • • • • • • • • • • • • • • • • • • • • • PARTICLE FILTER TR.

  21. PFT ISSUES •  •  •  •  Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 21

  22. APPEARANCE CHANGES • Same Color Objects • Background Clutter • Illumination Change • Shadows, Shades Use Depth! Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 22

  23. MODEL DRIFT PROBLEM • Templates Corrupted! Handle Occlusion! (No Model Update During Them) Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 23

  24. DEFICIENT FEATURE SPACE • * Local Optima of Feature Space • * Feature Noise • * Feature Failures Regularization Non-zero Values Normalization Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 24

  25. PERSISTENT OCCLUSION • Particles Converge to Local Optima / Remains The Same Region Advanced Motion Models (not always feasible) Restart Tracking (slow occlusion recovery) Expand Search Area! Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 25

  26. DYNAMICS… • * The Search is not Directed • * Neither of the Channels have Useful Information • * Particles Should Scatter Away from Last Known Position Occlusion! Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 26

  27. OCCLUSION Update model for target  Type of Occlusion is Important  Keep memoryvs. Keep focus on the target GENERATIVE MODELS DISCRIMINATIVE MODELS Combine Them! • do not address occlusion explicitly • maintain a large set of hypotheses • computationally expensive • direct occlusion detection • robust against partial & temp occ. • persistent occ. hinder tracking

  28. OCCLUSION TYPES • PTO partial occlusion • SAO self- or articulation occlusion • TFOtemporal full occlusion - shorter than 3 frames • PFO persistent full occlusion • CPOcomplex partial occlusion - including “split and merge” and permanent changes in a key attribute of a part of target • CFO complex full occlusion Kourosh Meshgi– ISHII LAB– MAY 2014 – Slide 28

  29. LITERATURE REVIW [Zhao & Nevatia, 04] Occlusion Indicator: Ratio of FG/BKG [Wu & Nevatia, 07] Handle Occlusion using Appearance Model [de Villiers et al., 12] Switch Tracker in the case of Occlusion [Song & Xiao, 13] Occlusion Indicator: New Peak in HOD or Reduction of the Size of Main Peak Many other papers handle occlusions as the by-product of their novel trackers

  30. SOLUTION • • • • • • • • • • • • • • • • • • • • • • • • OCCLUSION AWARE PFT

  31. Initialization Motion Model Observation Occlusion Flag? Calculate Likelihood NO YES Constant Likelihood Target Estimation PROPOSED MODIFICATION Occlusion Estimation Occlusion Threshold ? > Model Update NO YES Resampling Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 31

  32. OCCLUSION AWARE PARTICLE FILTER FRAMEWORK • Occlusion Flag (for each particle) • Observation Model • No-Occlusion Particles  Same as Before • Occlusion-Flagged Particles  Uniform Distribution Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 32

  33. TARGET ESTIMATION 1 • Position Estimation of the Target • Occlusion State for the Next Box x 0 1 x a 0 Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 33

  34. UPDATE RULE • Model Update (Separately for each Feature) • Modified Dynamics Model of Particle Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 34

  35. OA-PF DYNAMICS Occlusion! Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 35

  36. OA-PF DYNAMICS GOTCHA! Occlusion! Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 36

  37. OA-PF DYNAMICS • Quick Occlusion Recovery •  Low CPE • No Template Corruption • No Attraction to other Object/ Background Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 37

  38. FEATURES • COLOR (HOC) • TEXTURE (LBP) • GRADIENT (HOG) • EDGE (LOG) • DEPTH (HOD) • 2D PROJ. (BETA) • 3D SHAPE (PCL Σ) Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 38

  39. RESULTS • • • • • • • • • • • • • • • • • • • • • • • • & DISCUSSION

  40. DATASET( ) • Princeton Tracking Dataset 5 Validation Video with Ground Truth 95 Evaluation Video Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 40

  41. EXPERIMENT Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 41

  42. CRITERIA I • PASCAL VOC: Overall Performance 1 Success Area Under Curve to 0 1 Overlap Threshold Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 42

  43. RESULTS Success Plot 1 1 Success Rate 1 Overlap Threshold Kourosh Meshgi– ISHII LAB - MAR 2014- Slide 43 1

  44. CRITERIA II • Mean Central Point Error: Localization Success • Mean Scale Adaption Error Estimated Ground Truth Kourosh Meshgi– ISHII LAB– MAY 2014 – Slide 44

  45. RESULTS Center Positioning Error 400 CPE (pixels) 50 Frames

  46. RESULTS Scale Adaptation Error 140 SAE (pixels) 50 Frames

  47. CRITERIA III • FPhappens when a tracker doesn’t realize that the target is occluded. • MIhappens when the target is visible but the tracker fails to track it as if the target is still in an occlusion state • MT the estimated bounding box has nothing in common with ground truth box • FPS execution time in frames per second Kourosh Meshgi– ISHII LAB– MAY 2014 – Slide 47

  48. RESULTS Kourosh Meshgi– ISHII LAB– MAR 2014 – Slide 48

  49. PUBLICATION FINAL REVIEW Kourosh Meshgi– ISHII LAB– MAY 2014 – Slide 49

  50. FUTURE WORKS Kourosh Meshgi– ISHII LAB–MAY 2014 – Slide 50

More Related